search

openai / retro

Retro Games in Gym
MIT License
3.39k stars 526 forks source link

Emulator Performance #172

Closed christopherhesse closed 5 years ago

christopherhesse commented 5 years ago

It looks like inlining a method and creating some constants improves speed dramatically on ALE: https://github.com/mgbellemare/Arcade-Learning-Environment/pull/265

It's possible there are similar simple changes that could be made to other emulators.

endrift commented 5 years ago

Most of the other emulators have already had deep amounts of optimization already worked on, so Stella may be the odd one out here, not the average case.

mgbellemare commented 5 years ago

As FYI, Xitari (https://github.com/deepmind/xitari) is sped up about 100% relative to ALE, although it also stripped some actual functionality in the process...

endrift commented 5 years ago

Writing new libretro bindings for ALE's Stella might be a better idea than speeding up the newer version of Stella.

christopherhesse commented 5 years ago

Going to close this in favor of https://github.com/openai/retro/issues/127 since it sounds like most cores are already optimized a decent amount.

christopherhesse commented 5 years ago

Although with some fancy customizations, it's likely we still could improve performance of the emulators since we care more about speed than accuracy.